Test between fixed effects and random effects with clustered standard errors

[Lucca Horta]

Hello,
I'm testing the effects of salary dispersion across NFL teams using pooled model, fixed and random effects with clustered standard errors. So I tried the Hausman test, but I received the message "hausman cannot be used with vce(robust), vce(cluster cvar), or p-weighted data". So I would like to know how can I test between fixed effects versus pooled; fixed effects versus random effects; and random effects versus pooled. When not using clustered standard errors, I used the F test that appears on fixed effects regression, hausman and Breusch-Pagan lagrangean multiplier,
respectively. But with clustered errors the only test of those ones that worked well was the Breusch-Pagan lagrangean multiplier. Which tests can I do with clustered errors?
Thank you!

[Lucca Horta]

My regressions: xtreg colocacao gini eficpgto folha numjog idademed ntec nfc dnorte dsul dleste injury d14 d15 d16 d17 d18, re cluster(timeid) ["time" means "team" in Portuguese];
xtreg colocacao gini eficpgto folha numjog idademed ntec injury d14 d15 d16 d17 d18, fe cluster(timeid); and
reg colocacao gini folha eficpgto numjog idademed ntec nfc injury d18 d17 d16 d15 d14 dnorte dsul dleste, cluster(timeid).

[Carlo Lazzaro]

Lucca:
welcome to this forum.
I'm not clear with your using pooled OLS when you can easily test via -xttest0- if a random effect does exist.
That said, if -xttest0- gives you a statistical significant outcome, you can test if -re- specification (with non-default standard error) is actually the way to go for your panel datas set via the community-contributed programme -xtoverid- that you can spot and install typing from within Stata:

search xtoverid

If the -xtoverid- outcome reaches statistical significance, you should go -fe-.
If the -xtest0- outcome does not allow you to reject the null, you should switch to pooled OLS.

[Lucca Horta]

Thank you very much for your answer. It helped a lot!
Lucca Horta.

[Phil Bromiley]

Note that the fixed effect specification will estimate the effect of deviations in salary dispersion from the team mean which may not be the same as the effects of the general level of salary dispersion for a team. That is, the fe and re parameters might differ either because the true parameters on the between differ from the within, or because correlation between the omitted panel level variables and included variables makes re inconsistent. Is the effect of a team increasing dispersion by 1 the same as the effect of one team having a general level of dispersion 1 higher than another? Given that dispersion probably fits into many features of the team, deviations and with-in levels could easily have different true effects.

Whether the right model is to look at within deviations from the firm mean or to look at levels in a between or xthybrid model is partly a matter of your theory and understanding of the problem.